Riccardo Crescenzi, Andrés Rodríguez-Pose, and Michael Storper The territorial dynamics of innovation in China and India

Transcription

1 Riccardo Crescenzi, Andrés Rodríguez-Pose, and Michael Storper The territorial dynamics of innovation in China and India Article (Accepted version) (Refereed) Original citation: (Crescenzi, Riccardo, Rodríguez-Pose, Andrés and Storper, Michael (2012) The territorial dynamics of innovation in China and India. Journal of Economic Geography, 12 (5). pp ISSN DOI: /jeg/lbs The Authors This version available at: Available in LSE Research Online: November 2014 LSE has developed LSE Research Online so that users may access research output of the School. Copyright and Moral Rights for the papers on this site are retained by the individual authors and/or other copyright owners. Users may download and/or print one copy of any article(s) in LSE Research Online to facilitate their private study or for non-commercial research. You may not engage in further distribution of the material or use it for any profit-making activities or any commercial gain. You may freely distribute the URL ( of the LSE Research Online website. This document is the author s final accepted version of the journal article. There may be differences between this version and the published version. You are advised to consult the publisher s version if you wish to cite from it.

2 THE TERRITORIAL DYNAMICS OF INNOVATION IN CHINA AND INDIA Riccardo Crescenzi*, Andrés Rodríguez-Pose**, and Michael Storper*** Final Accepted Version Please cite as: Crescenzi, Riccardo, Rodríguez-Pose, Andrés and Storper, Michael (2012) The territorial dynamics of innovation in China and India.Journal of economic geography, 12 (5). pp ISSN doi: /jeg/lbs Abstract: This paper analyses the geography of innovation in China and India. Using a tailor-made panel database for regions in these two countries, we show that both countries exhibit increasingly strong polarisation of innovative capacity in a limited number of urban areas. But the factors behind this polarisation and the strong contrasts in innovative capacity between the provinces and states within both countries are quite different. In China, the concentration of innovation is fundamentally driven by agglomeration forces, linked to population, industrial specialisation and infrastructure endowment. Innovative regions in China generate few knowledge spillovers to other regions, and create strong backwash effects to leading centres. In India, by contrast, innovation is much more dependent on a combination of good local socioeconomic structures and investment in science and technology. Indian innovation hubs also generate positive knowledge spillovers to other regions. Keywords: Innovation, R&D, socioeconomic conditions, geography, regions, China, India. JEL classifications: R11, R12, O32, O33 * Riccardo Crescenzi London School of Economics r.crescenzi@lse.ac.uk ** Andrés Rodríguez-Pose London School of Economics & IMDEA Social Science - A.Rodriguez-Pose@lse.ac.uk *** Michael Storper - London School of Economics & Sciences Po M.Storper@lse.ac.uk 1

3 Acknowledgements The authors would like to thank ChinChih Chen and Max Nathan for their excellent research assistance. Financial support by ESPON KIT (European Observation Network of Territorial Development and Cohesion - Knowledge, Innovation, Territory) Project and the LSE-SERC, as well as by the European Research Council under the European Union s Seventh Framework Programme (FP7/ )/ERC grant agreement nº and of a Leverhulme Trust Major Research Fellowship, is gratefully acknowledged. The authors are also grateful to participants in the seminars held in Amsterdam, Barcelona, Bergamo, Milan, Paris and Rome for comments on earlier drafts of this paper. The authors are solely responsible for any errors contained in the article. 2

4 1. INTRODUCTION The world is witnessing a significant change in the geography of innovation. Until recently, research and development (R&D)-led innovation was heavily concentrated in the countries of the so-called Triad. The United States (US) and Canada, the European Union (EU) and Japan did most of the investment in R&D and, consequently, generated most of the innovation and, in particular, of the radical innovation (Dosi et al., 2006; Furman and Hayes, 2004). World technology gaps overall are not disappearing, but because the world economy is becoming generally more innovative, any successful developer is required to be more innovative than at points in the past (Kemeny, 2011). Moreover, there is some turbulence in the ranks of developing countries, with a few successful countries moving up world technology ranks, little-by-little. Thus, the privileged position by the countries of the Triad is increasingly being challenged by the emergence of innovation hubs in a few emerging economies The so-called BRICS countries (Brazil, Russia, India, China and South Africa) are perhaps beginning to move closer to the Triad s leadership position. Largely reflecting their efforts, R&D investment has become more globalised (Lundvall et al., 2009; Fu and Soete, 2010) and certain emerging countries in Asia and elsewhere are increasingly basing their economic growth on their capacity to generate new product and process innovations (Mahmood and Singh, 2003; Popkin and Iyengar, 2007). Throughout most of the 1990s, export growth in China and India was based principally on their cost advantage, and Chinese and Indian firms acted either as production platforms for Western firms or attempted to pursue indigenous innovation strategies (Yeung, 2009). Since the turn of the century and complementing their cost advantage Chinese and Indian firms have been moving up the value chain. To do so, they rely on R&D-led innovation and international partnerships (Bruche, 2009; Kuchiki and Tsuji, 2010). Chinese and Indian firms owe their innovation in part to the development and growing maturity of the local innovation systems in those countries (Lundvall et al., 2009). The rapidly changing roles of China and India in the global innovation picture has attracted considerable attention, from many different explanatory perspectives (Popkin and Iyengar, 2007; Parayil and D Costa, 2009). Santangelo (2005) and Fu and Soete (2010), for example, have analysed how these countries are catching up on technology. Malerba and Mani (2009) have concentrated on the sectoral dimension of the 3

5 technological catch-up, while Lundvall et al (2009) and Kuchiki and Tsuji (2010) have focused, respectively, on the emergence of innovation systems and clusters. Most such analyses of innovation in China and India emphasise changes in the conditions for the global spread of innovation, or national factors. By contrast, there is little literature on the ways that subnational processes of matching capital, labour and knowledge might affect national innovation output in the two countries. Indeed, beyond a handful of studies, 1 relatively little is known about the geography of innovation across Chinese and Indian provinces and states and virtually nothing about which factors, conditions, and interactions determine why some areas in these countries are more innovative than others (e.g. da Motta e Alburquerque 2003; Sun, 2003; Mitra, 2007; Fu, 2008; Wang and Lin, 2008). A certain number of studies adapt the regional innovation system approach to developing country contexts (e.g., Asheim and Vang, 2004; Chaminade and Vang 2008; Scott and Garofoli, 2007; Padilla-Pérez et al, 2009) These studies emphasise the importance of trans-national processes and global-local interactions in shaping regional outcomes especially the role of multinational enterprises (MNEs) and their links with local firms (Schmitz, 2007; Yeung, 2009), the multiple contributions of diaspora communities (Saxenian, 2006; Filatotchev at al. 2011). They also stress the role of country-specific institutional and policy factors in mediating these global forces (Yeung, 2006) and supporting the upgrade of indigenous human capital and local SMEs (Chaminade and Vang 2008). However, the great majority of this literature relies on detailed case-studies. They tend to focus on the most successful and interesting cityregions. We therefore lack more systematic, large-scale quantitative evidence, especially cross-country and panel data studies (Padilla-Perez et al., 2009), which could help corroborate whether the insights from case-study analyses are broadly applicable, and which could shed insight into whether innovative regions emerge for the same reasons in the two countries. 1 Liu and Sun (2009), using invention patent data, provide a systematic overview of the evolution of the patenting activity of Chinese provinces between 1985 and 2005 and compare it to that of US states. The paper unveils a number of differences between the geography of innovation in China and in the US, but falls short of analysing the factors which determine the geographical differences between the two countries. It also suggests that the regional innovation policy of the Chinese government should follow a so-called basic law of the spatial distribution of innovative activities, meaning a greater focus on innovation in coastal areas. 4

6 In order to address these issues, this paper presents a systematic, cross-country quantitative analysis of the geography of innovation in China and India, arguing that each country has a distinctive geographical process of bringing together the capital, labour and knowledge that underlie innovation. It deploys new panel data sets to explore this geographical matching process. In order to do this, we first briefly examine some stylised facts about the geographies of innovation in China and India. We then present a model aimed at identifying the factors behind the innovative capacity of Chinese provinces and Indian states. In section 4 we introduce the results of the analysis, highlighting the important differences in the geographies of innovation between China and India. The final section presents the main conclusions and some preliminary policy implications. 2. GEOGRAPHIES OF INNOVATION IN CHINA AND INDIA: STYLISED FACTS As in most parts of the world, innovation, measured by patent applications, in China and India is unevenly spread geographically. In China, innovative activity is highly concentrated along the eastern seaboard, and especially in the South and in the two largest cities (Sun, 2003; Wang and Lin, 2008). Guangdong, with 46% of total patenting activity 2 is the leading province. Guangdong is followed at some distance by Beijing (14%) and Shanghai (13%). The top three Chinese regions account for 73% of all patents. Most other coastal provinces have shares of patent applications which range between 1 and 3% of the total. The Centre and the West of the country, despite counting with some populous provinces, are far less innovative. Only Sichuan (South West) and Hunan (Centre) have more than 1% of all Chinese patents, while western provinces, such as Tibet and Qinghai, and some central provinces, such as Ningxia, barely generate any patenting activity. The geography of patenting in India is dominated by the high-tech hubs of the country. Cities such as Bangalore, Chennai, Delhi, Hyderabad, Mumbai and Pune are at the origin of the great majority of patents (Mitra 2007). At the regional level, da Motta e Alburquerque (2003) find that from 1981 until 2002, nearly half of all patents were assigned to Indian inventors in two states, Maharashtra and Delhi. Our data echoes this; Maharashtra (capital Mumbai) and Delhi respectively account for 26% and 24% of total patents. Andhra Pradesh, with the great majority of its 13% of Indian patents centred in 2 Patenting activity as measured by patents filed under the Patent Co-operation Treaty (PCT). 5

7 its capital Hyderabad, comes third. All together, the three top Indian states account for 64% of all patents. Other states around Delhi or in the South, such as Karnataka (8.7%, South, capital Bangalore), Haryana (7%, close to Delhi) and Tamil Nadu (7%, South, capital Chennai) form a second tier of innovative states. In contrast, states in the North East are less innovative. This less innovative group includes Uttar Pradesh and West Bengal, the first and fourth most populous states in India. And there is no recorded patent activity until 2007 for most of the north-eastern border states, including some relatively big states such as Assam (with a population of 31 million, according to the 2011 census). Figure 1 illustrates the distribution of patenting across space in China and India in 2007, focusing on the 20 regions with the highest patent counts. The EU and USA are included in the Figure for comparison. Figure 1 - Spatial distribution of patenting: Top 20 most innovative regions, 2007 (China, India, EU and US) Cumulative percentage (Total Patents, 2007) Region's Ranking (less innovative to most innovative) Source: OECD, 2010 Note: China: 31 provinces Cumulative distribution in China, 2007 India: 24 states Cumulative distribution in India, 2007 USA: 179 BEA Economic Areas EU24 (Cyprus, Lithuania and M alta not included): 841 OECD TL3 regions Cumulative distribution in USA, 2007 Cumulative distribution in EU24, 2007 What stands out in Figure 1 is the clear contrast between the spatial structure of mature and emerging innovation systems. In mature innovation systems, such as 6

8 those of the EU or the US, despite a high degree of geographical agglomeration, patenting activity is spread across a greater number of regions than in emerging systems, where patenting is heavily concentrated in the top three, in the case of China, or in the top six, in the case of India, regions. Whereas in Europe the 20 most innovative regions account for roughly 50% of the patents, this threshold is reached by a single province in China (Guangdong) and by three Indian states (Maharashtra, Delhi and Andhra Pradesh). However, there are also significant differences between the level of concentration of innovative activities in the two mature and the two emerging innovation systems considered. In the mature systems and as mentioned by earlier research (Dosi et al., 2006; Crescenzi et al., 2007), patenting is more territorially concentrated in the US than in the EU. 3 In the emerging systems, the six highestpatenting regions in China account for a bigger share of innovative activity than those in India, although the pattern reverses after that with a long tail of Indian regions. Over time the trend both in China and India has been towards a rapid increase in patenting, accompanied by a rise in the geographical concentration of innovative activity (Sun, 2003). In 1994 patenting in India was far more concentrated than in China. However, the situation reversed right after 2000 when innovation became more agglomerated in Chinese provinces than in Indian states. Since then the trend has been towards an ever greater concentration of innovation in a few coastal provinces (Liu and Sun, 2009) accompanied by significant structural changes in local-level systems of innovation conditions (Su and Liu, 2010). These changes emphasize the cumulative nature of the agglomeration process. Such polarisation is further reinforced by an emerging trend towards territorial competition among Chinese local authorities for the attraction of external resources from both the central government and international investors (Chien and Gordon, 2008). What are the reasons behind the rapid expansion of patenting in both China and India and behind the territorial agglomeration of patenting activity? Both India and, in particular, China have invested heavily in innovation inputs. Both countries have witnessed rapidly rising literacy rates and higher education enrolment: in 2007 China had 25m university students and India 13m, in comparison to 12m in the US (Dahlman, 3 As far as the EU-US comparison is concerned, the magnitude of this gap is partially attenuated when OECD TL2 (larger) Regions are considered for the EU. However, OECD TL2 regions correspond to US States while TL3 to Economic Areas. The evidence on the spatial concentration of innovative activity remains qualitatively unchanged, irrespective of the spatial scale. 7

9 2010). Moreover, the rise in university placements in these two countries has been absolutely phenomenal. Since 1999, China has seen an annual growth in student numbers of 20 per cent or more (Schaaper, 2009). Similarly, during the 1990s Indian universities significantly increased their output of engineering graduates from 44,000 per year in 1992 to 184,000 in This compares with 352,000 per year in China and just 76,000 per year in the US (Mitra, 2007). China has also exploited global knowledge by moving students abroad to study in larger numbers than India and Indian returnees have had significant impacts on the country s ICT sector (Saxenian, 2006). With 926 R&D researchers per million people, China now has the second-highest total number of researchers world-wide (Schaaper, 2009). Expenditure in R&D has been at the forefront of the innovative effort in China. China s overall R&D spending as a share of GDP almost trebled between 1995 and China lagged behind India for most of the 1990s in overall R&D expenditure, but overtook its neighbour at the end of the decade and has continued to widen the gap ever since. By contrast, India s expenditure in R&D relative to GDP remained stable during the same period. Moreover, India s headline figure conceals a large R&D rise in pharmaceuticals, ICT, electronics and auto parts (Dahlman, 2010). Both figures also hide important differences in institutional capacity to invest. The Indian government s science and technology spending exhibits a much greater year-on-year volatility than that of China. How have these investments influenced innovation outputs? Despite an increase in relative investment in science and technology projects in the 1970s and 1980s, China and India exhibited low patent counts in these decades, with a jump in patenting activity finally appearing in the mid-1990s. Figure 2 shows this jump in national innovation performance as measured by patent applications per million people. Both countries have raised their innovation rates, as measured by patents, particularly in the last decade, but China much more so than India. 8

10 Figure 2 Evolution of patent intensity over time in China, India, EU and the USA, Patents per million inhabitants Primary Axis EU and USA Evolution of Patent Intensity, Patents per capita in China, India, EU and USA ( ) Secondary Axis (China & India) 0 Source: OECD PatStat USA EU27 EU15 China India 0 Indian patenting has increased steadily since the 1990s, while Chinese patenting accelerated from 2005 onwards. Schaaper (2009) reports a huge surge in patent applications at the Chinese patent office and large increases in international patenting by Chinese inventors and firms. Both India and China still have some way to go before reaching US or EU levels of patenting activity (Mahmood and Singh, 2003; Schaaper, 2009), as well as those of other Asian Tiger economies (Tseng, 2009). China and India also lag behind some neighbouring countries. Tseng (2009) compares Chinese and Indian patenting in Information and Communication Technologies (ICT) to that in other Asian economies. During the period , the largest number of patent granted in ICT was in South Korea (22,612 in total), followed by Taiwan (19,907), Singapore (1,333), and Hong Kong (622). As an average for the period, China and India record only 440 and 81 ICT patents granted in total respectively. Social and institutional factors account for a great deal of the differences observed in recent innovation trends between both countries. Historically, neither China, nor India have been strangers to the use of innovation and technology-led development policies in order to pursue national prestige or increase their international geopolitical standing. 9

11 Their space flight and atomic weapons programmes are clear examples of this type of top-down science and technology policies, more aimed at achieving immediate symbolic returns than at enhancing the broad performance of their national innovation systems (Leadbeater and Wilsdon, 2007). Nonetheless,, there is a drive to move away from heavily dominated statist models of public policy and towards market-led reforms (Jian et al, 1996; Fan, 2008; Fleischer et al, 2010; Su and Liu, 2010). China s earlier move to globalise its economy and revamp its innovation system has borne fruit and is one of the reasons behind the growing gap in innovation between China and India. China s greater emphasis on trade, FDI or the licensing of foreign technology, among others, has produced significant returns in innovation outputs (Dahlman, 2010). India's reform and pace of change, despite taking off after 1991, has been dragged down by a large bureaucracy and less malleable institutions. 3. MODEL AND DATA 3.1. The structure of the model In order to explain these trends and the differences in innovation geographies between China and India, we draw on the three major theoretical strands on the causes of innovation and its geography to build a modified regional knowledge production function (Crescenzi et al. 2007; O huallachain and Leslie, 2007) inspired by the traditional framework of Griliches (1979 and 1986) and Jaffe (1986). In this production function, regional innovation is proxied by regional patent intensity and explained by a series of factors. First, following endogenous growth models, innovation may be a consequence of the returns of investment in human capital and innovation inputs, such as R&D spending, which lead to higher patenting rates. China and India partly adopted this sort of approach and continue to invest heavily in innovation inputs, such as R&D and higher education (Kuijs and Wang, 2006). The economic geography approach to the genesis of innovation holds that differences in innovative performance across territories will emerge from agglomeration economies, because concentrations of firms and skilled workers will increase the creation and diffusion of knowledge, or what are known as localised spillovers or learning (e.g. Acs et al, 2002; Carlino et al, 2007). Finally, the literature on regional innovation systems has emphasised regional-level factors, which 10

12 include universities and public agencies, networks (e.g. public-private partnerships) and local institutions (Cooke, 2002). How these factors combine in space makes the geographic configuration of economic agents (...) fundamentally important in shaping the innovative capabilities of firms and industries (Asheim and Gertler, 2005: ). We further extend this traditional framework by accounting for the role of territorial characteristics and spatial processes (Audretsch and Feldman, 1996; Crescenzi et al., 2007; O huallachain and Leslie, 2007; Ponds et al, 2010). In this way, we are able to consider both systems of innovation conditions and other internal and external factors. The model takes the following form: y α τ β γ δ ζ ϑ + ε i, t = i + t + R& Di, t + WR& Di, t + SFi, t + WSFi, t + xi, t i, t (1) where: y R&D SF WR&D and WSF x ε represents Regional Patent intensity; is the share of R&D/S&T Expenditure in regional GDP; is the Social Filter Index; are spatial lags of R&D/S&T and SF respectively with appropriate spatial weights; is a set of structural features/determinants of innovation of region i; is an idiosyncratic error; and where i represents the region and t time. The choice of empirical variables included in the model is set out in Table 1. Table 1. Empirical variables included in the model. Variable Internal Factors External Factors R&D Local Investment in S&T/R&D Investment in S&T/R&D in neighbouring areas Social filter Structural characteristics Same characteristics in 11

13 Specialisation Relative wealth Agglomeration economies Infrastructure endowment Mobility of people Fixed effects that would make a region neighbouring areas more innovation prone, including: Human Capital Sectoral composition Use of resources (unemployment) Demographics Krugman Index GDP per capita Population Density Kilometres (Kms) of motorways/railways Migration rate Region/Province-specific fixed effect + Time Trends Two panel datasets for Chinese provinces and Indian states are assembled. Data for China cover 30 provinces from 1995 through Data for India cover 19 states between In the following section we describe the variables included in the model in detail (see Appendix A for technical specifications and data sources) Variables included in the model Dependent variable Our dependent variable is patent intensity. Patent intensity is measured by the number of regional patents per capita and is used as a proxy for the innovative performance of the local economy. 4 For China data are available for the Provincial-level administrative subdivisions: 22 Provinces, 4 Autonomous Regions, 4 Municipalities. Two Special Administrative Regions (Hong Kong and Macau) and One Autonomous Region (Tibet) have been excluded from the analysis due to the lack of data for the selected variables. 5 For India data are available for 18 States and 3 Union Territories. Bihar and Rajasthan are included in descriptive statistics but not in the regression analysis due to the limited number of observations available over time. 12

14 The use of patent intensity as our measure of regional innovation in China and India is not without controversy. Not all industries tend to patent with the same intensity and, in an emerging country context, the number of firms or organisations in a position to file patterns is rather limited. However, there are no reliable substitute measures and other readily available innovation metrics for India and China tend to follow spatial patterns similar to patents. For instance, multinational firms location patterns closely match those of patents: between 60-80% of all MNEs are concentrated in the Beijing- Shanghai-Guangdong axis, in the case of China, and in Bangalore/Pune/National Capital Region, in India (Bruche, 2009). We also use patent data from Patent Cooperation Treaty (PCT) patent applications, 6 avoiding some much-discussed issues with domestic patent coverage and quality, especially in China (Li and Pai, 2010). A PCT filing can be seen as a worldwide patent application and is much less biased than national applications. ( ) Further, the PCT reflects the technological activities of emerging countries quite well (Brazil, Russia, China, India, etc.). (OECD 2009: 66) Two important caveats remain. First, patents measure invention and tend to be biased towards particular sectors of the economy where inventions are protected via patenting (rather than via trademarks or secrecy, for example) (OECD, 2009). Fagerberg and Shrolec (2009) argue that since minor innovations/adaptations will not be patentable most innovative activity in emerging countries will not be counted. This is an important issue in the case of China and India, although as both countries rapidly approach the technological frontier in a number of sectors, this objection may be becoming less pertinent than before. Second, patent counts include both domestic and foreign firms, meaning that they may not fully capture domestic innovation capacity (Li and Pai, 2010; Wadhwa, 2010). In the case of MNEs, patents may be filed in any office around the world, regardless of where the invention actually takes place, making it hard to assign patents to specific territories. Duan and Kong (2008), in a study of Chinese patents , observe that most Chinese applications to the USPTO are owned by foreign firms. In India the picture is more complex. Da Motta e Alburquerque (2003) finds that between 1981 and 2001, a third of Indian patent applications to the USPTO had foreign assignees, higher than 6 The Patent Cooperation Treaty (PCT) was concluded in 1970 ( ) and makes it possible to seek patent protection for an invention simultaneously in each of a large number of countries by filing an international patent application. World Intellectual Property Organisation (WIPO) website l (accessed on May 18th 2012) 13

15 Brazil, South Africa or Mexico. However, Mani (2004) examines Indian-based inventors during the 1990s, reporting that 85% of the patents granted were awarded to public research institutes, as well as some to local firms and individuals as opposed to foreign affiliates located in India. Independent variables As indicated in model (1), the independent variables cover internal conditions spillover-related factors, and a set of other structural conditions potentially affecting the geography of innovation in the two countries (GDP per capita, transport infrastructure, agglomeration, migration flows). The internal conditions include R&D expenditure and the social filter. R&D expenditure is measured by the percentage of regional GDP devoted to S&T (China) or R&D (India). This indicator has been frequently used in the literature as a measure of the allocation of resources to research and other information-generating activities in response to perceived profit opportunities (Grossman and Helpman, 1991: 6), as well as as a proxy for the local capability to absorb innovation produced elsewhere (Cohen and Levinthal, 1990; Maurseth and Verspagen, 2002). The social filter aims at capturing the structural preconditions for the successful development of regional innovation systems. As underlined by the innovation systems approach, the capacity of any given region to generate and use knowledge depends on a complex set of local factors. The most common of these factors encompass regional social and business networks; social stratification; and levels of modernity versus tradition. While such factors can be relatively easily identified in case-studies, they must be captured in a more parsimonious way for cross-sectional analysis by a combination of innovative and conservative elements that favour or deter the development of successful regional innovation systems across a wide variety of places (Rodríguez-Pose, 1999: 82). The social filter variable used in this paper is therefore made up of a set of variables available for both China and India in a consistent and comparable fashion, focusing on three main aspects of an innovation prone social structure: educational achievement (Lundvall, 1992; Malecki, 1997); the productive use of human resources (Gordon, 2001); and demographic structure and dynamism (Rodríguez-Pose, 1999). The combination of proxies for all these different dimensions into one single composite indicator (the Social Filter Index) develops a quantitative 14

16 profile of an innovation prone regional environment, making it possible to compare the social filter conditions of different regions across countries (Crescenzi and Rodríguez- Pose, 2011). The first domain is human capital attainment, expressed by the shares of adult population who have completed tertiary education. We expect the stock of human capital, an innovation input, to be positively related to the rate of innovation. The second domain, the structure of productive resources, is measured by the percentage of the labour force employed in agriculture and the rate of unemployment. Both should have a negative association with innovation. Over the past two decades, India and China have been experiencing both large scale rural-urban migration and industrialisation, factors linked to improved innovative performance, and a declining salience of agricultural activity (Dahlman, 2010; Gajwani, et al., 2006). Higher long-term unemployment rates indicate weak local labour demand, and also suggest poor quality human capital (as opposed to education-based quantity measures) (Gordon, 2001). For the third domain, we use the percentage of population aged between 15 and 24 for the flow of labour entering the labour force, potentially refreshing the existing stock of knowledge and skills (Crescenzi et al., 2007). We fit the social filter both as a set of individual variables, and as a social filter index constructed through Principal Component Analysis (PCA). The PCA output is shown in Tables B-1 and B-2 in Appendix B. The first principal component alone accounts for 45 and 36 percent of the total variance of the original variables considered for China and India respectively. The scores are computed from the standardised value of the original variables by using the coefficients listed under Comp1 in Table B-2, generating the social filter index. Both R&D expenditures and the social filter can have effects beyond the borders of regions. Therefore, in addition to describing the internal characteristics of each territory, the model also includes variables representing the characteristics of neighbouring regions that may influence the innovative performance in the region of interest. The potential of R&D expenditure to spill over beyond regional borders is depicted by the variable WR&D. This spatially lagged R&D variable captures extra-regional innovation, by measuring the aggregate impact of innovative activities pursued in 15

17 neighbouring regions, as they could exert a positive impact on local innovative performance, via inter-regional knowledge exchange channels and complementarities. Conversely, centripetal forces driving the location of innovative activities toward preexisting hot-spots may lead to the generation of negative externalities: proximity to innovative areas may drain resources from nearby areas. Following Rodríguez-Pose and Crescenzi (2008), the extra-regional innovative activity is proxied by the average of R&D/S&T intensity in neighbouring regions and is calculated as: n WR & D = R & D w with i j (2) i j = 1 j ij where R&D is our proxy for regional innovative inputs of the j-th region and wijis a generic spatial weight. In order to test for the spatial scope of the processes discussed above, alternative definitions for the spatial weights have been adopted in our analysis: highly localised spatial processes have been proxied by means of first-order contiguity weights (w FC ), while long-distance flows have been captured with inverse-distance weights (w ID ): 7 FC 1 if j directly shares a border or a vertex with i w ij = (3) 0 otherwise 0 if i = j 1 w ID ij = dij (4) if i j 1 j dij where d ij is the linear straight-line distance between region i and j and w the corresponding weight. 7 Alternative definitions for the spatial weights matrix are possible: distance weights matrices (defining the elements as the inverse of the distances) and other binary matrices (rook and queen contiguity matrices). 16

18 The same principle as for WR&D is applied to the social filter. A variable WSF is created in order to capture potential extra-regional social filter spillovers. The measure of extraregional social filter conditions is calculated following the same principle presented in equation (2). For each region i: WSF i = n j = 1 SF w j ij with i j (5) where SF is our social filter index and w is as above. Finally, our third group of independent variables includes a vector of additional key drivers of innovation. These include economic specialization; levels of GDP per capita; infrastructure endowments; agglomeration levels; and inward migration. The degree of specialisation is measured, following Midelfart-Knarvik et al. (2002), by the Krugman Index, which is calculated according as follows: k a) for each region, the share of industry k in that region s total employment: ν (t); b) the share of the same industry in the employment of all other regions: ν i (t); and c) the absolute values of the difference between these shares, added over all industries: k i k k k k = k i with i ( t) = x ( t) / x ( t) j i i k j i i k K ( t) abs( ν ( t) ν i ( t)) i ν (6) The index takes the value zero if region i has an industrial structure identical to the rest of the country, and takes the maximum value of two if it has no industries in common with the rest of the country. The initial level of GDP per capita is introduced in the model in order to account for the region s initial wealth as proxy for the distance from the technological frontier (Fagerberg, 1994). The significance and magnitude of the coefficient associated to this variable allows us to test for the existence of technological catch-up. Transport infrastructure may affect innovative performance through a variety of mechanisms. In order to capture the direct impact of transport infrastructure on 17

19 regional growth, the model includes a specific proxy for the stock of transport infrastructure, measured by the total motorways or railways in the region, in kilometres, standardised by the regional population (Canning and Pedroni, 2004). See Table A-1 in the Appendix for further detail. Another independent variable included in the model is agglomeration. As indicated by the new economic geography approach, the geographical concentration of economic activity has an independent impact on innovation (Duranton and Puga, 2003; Charlot and Duranton, 2004) and thus needs to be controlled for in order to single out the impact of other knowledge assets such us R&D intensity and Social Filter conditions. We use population density as our proxy of agglomeration. The regional rate of migration (i.e., net inflow of people from other regions) 8 is also included in the model in order to measure the capacity of the region to benefit from external human capital and knowledge by attracting new workers, increasing the size of its labour pool and its diversity in terms of skills and cultural background (Ottaviano and Peri, 2005). Finally, in the case of China we include the share of state-owned (SOEs) industrial firms as a percentage of all industrial firms. State ownership may affect firms propensity to innovate and the technological opportunities available to them, as well as the likelihood of forming partnerships with MNEs (Li et al., 2007; Motohashi and Yun, 2007; Gu et al., 2007). SOEs are likely to have politically-motivated locations, and possibly to be less innovation-oriented than privately-owned firms. 4. RESULTS OF THE EMPIRICAL ANALYSIS Using these variables, we estimate model (1) as a two-way fixed effects panel data regression. 9 In order to minimise any potential spatial autocorrelation, we explicitly 8 For China, both the net and gross internal rate of migration are available. For India, only the gross in-flow of people into each state is available due to the lack of data on outflows. However in the case of China the correlation between net and gross migration rate is 0.95 and regression results are qualitatively identical if net migration is replaced by gross migration in order to match the variable definition of the dataset for India more precisely. 9 The Hausman test indicates that fixed effects is the preferred estimation, rejecting the random-effects specification. In addition the F-Test confirms that the region-specific effects are statistically significant. 18

20 control for national growth rates. Moreover, the introduction of the spatially lagged variables WR&D and WSF allows us to take into consideration the interactions between neighbouring regions, minimising any effect on the residuals. 10 The analysis also uses robust standard errors clustered by state (India) or province (China). We deal with potential endogeneity of the right-hand side variables by fitting these as one-period lags. Finally, because of different accounting units, we express all explanatory variables as a percentage of the respective GDP or population. This is an exploratory analysis aimed at uncovering the territorial dynamics of innovation in the two countries rather than identifying causal relationships consequently, in what follows, we focus mainly on the sign and significance of coefficients, rather than the size of specific point estimates. 11 The results are shown in Tables 2 and 3. Tables 2a and 3a give the main results for China and India respectively. In each case, models (1) through (3) explore the more traditional components deemed to affect innovation by regressing patenting rates on R&D/S&T expenditure and various spatial lags of science spending. Models (4) through (7) introduce the social filter and its spatially weighted variants. Models (8) through (12 for China, or 11 in the case of India) bring in the wider structural factors, as well as any country-specific variables. Tables 2b and 3b decompose the social filter index into its constituent elements respectively for China and India. The model generally performs better for Chinese data, as we have a longer time period and more observations (because of a larger number of spatial units for a longer time series). Results for India tend to be more volatile China 10 The absence of spatial correlation is confirmed by conducting Moran s I test for each year. The results of these tests are not significant for the majority of years. 11 For China the results presented in the tables are based on a fully balanced panel dataset with data for all provinces and years covered by the analysis. For India data limitations are more significant and make it impossible to produce an equally balanced panel dataset. While the results presented in the tables are based exclusively on the data available from the original sources a number of tests have been implemented in order to test the robustness of the results to different sample sizes. In order to increase the number of available observations and test the robustness of our results we used linear interpolation (i.e. new data points are constructed only within the range of a discrete set of known data points) of missing values for patent intensity only. This procedure increased the number of available observation with qualitatively unchanged results, confirming the robustness of our results. 19

21 The estimations for China are given in Table 2a. Regressions (1) to (3) explore the linear elements of the innovation process. The results indicate that regional R&D spending is not significantly connected to local patenting, and that spatially weighted science and technology spending is negative and insignificant. This echoes other findings for the European Union, where R&D spending tends to be centralised at nation state level generating a detachment between local innovative efforts and localised output (Crescenzi et al., 2007). The fact that R&D spillovers are also not significant (contrary to large part of the existing evidence on the EU) is also possibly a sign of a disconnect between the more dynamic innovation hubs in China and a national R&D policy which may still aim to counterbalance the concentration of innovative activities by funding R&D activities in more remote or less accessible regions for either strategic (i.e. keeping some military industry away from the coast) or development reasons. Table 2a around here The introduction of the social filter in regressions (4) through (8) leads to an interesting twist in the story. The social filter index is positively and significantly (at the 5% level) associated with innovation rates. However, the introduction of agglomeration indicators removes this significance. As in the case of the R&D spillovers, the spatial lags of the social filter (using first order contiguity weights) are negative, becoming negative and significant when structural factors are controlled for. Models (9) to (12) include a wider set of structural factors. The introduction of agglomeration measures dominates the analysis. Both the Krugman index and population density have a positive and significant connection to innovation at the 1% level. Railway density has a large point estimate but is only marginally significant in the full models, perhaps because China has shown a preference until recently for building roads. Net migration is positive and significant at the 1% level, but point estimates are much smaller than for agglomeration measures. To further explore agglomeration processes, we interact science and technology spending with population density. The interaction term is negative and significant at 5%, but renders local R&D spending positive and significant. Finally, model (12) includes a control for the share of state-owned firms. As predicted, we find a negative coefficient potentially indicating that a strong presence of state- 20

22 owned enterprises may drag down regional innovation although the coefficient is not significant. Table 2b around here The decomposition of the social filter into its constituent variables is presented in Table 2b. The results of this analysis corroborate those of Table 2a, highlighting the robustness of the exercise. Reflecting a weak association between the social filter index and innovation in China, none of the indicators making up the social filter index has a robust connection to innovation. While in the earlier regressions [Regressions(1) to (4)] the presence of a young population or the levels of human capital are positively and significantly associated with regional innovation and agricultural employment has a negative connection with it, the introduction of the indicators depicting levels of agglomeration, industrial specialisation, and infrastructure endowment in the analysis in regressions (5) to (9) renders these connections insignificant. By contrast, the relatively strong associations between spatially weighted S&T variable, on the one hand, and between agglomeration, industrial specialisation and transport infrastructure density, on the other, remain largely untouched. Overall, the results suggest that the geography of innovation in China is akin to what could have been predicted under a new economic geography framework. It is a traditional agglomeration story: richer regions with an intense agglomeration of activities, good infrastructure endowments, and a greater degree of industrial specialisation not only have higher patenting rates, but also absorb innovative potential from neighbouring areas. When agglomeration effects are taken into account, the R&D spillovers become negative and significant, generating what is known as the Krugman shadow effect (McCann and Ortega-Argilés, 2011), that is the agglomeration of innovation in core areas leads to ever greater concentration of innovation by promoting further outflows of knowledge from neighbouring regions. This drawing of resources from surrounding areas is a sign of the presence of what can be considered as a less mature innovation system. It might appear paradoxical that China has geography of innovation led by agglomeration forces, which are generally considered to be a strong feature of a marketdriven economy. The territorial distribution of Chinese innovation is characterised by the overwhelming concentration of innovative activity in Guangdong province. This 21

23 mega-agglomeration of export-oriented industry is the country s main innovation hub, but the Guangdong agglomeration is in many ways the result of a national strategy designed to turn China into the workshop of the world. However, this top-down strategy was subsequently allowed to interact with market forces in a powerful way. Other innovative areas have also benefited from policy intervention by the Chinese government. As Wang (2010) notes, the particular importance of Special Economic Zones (SEZs) in 1978, which acted to spatially concentrate FDI flows and thus technology transfer have played an important role in developing clusters of high-tech innovative activity in Southern and coastal regions such as Guangdong, Fujian, Hainan, Hunchun, and the Pudong Development Zone (Shanghai). Inland regions, on the other hand, were allowed far less integration with the outside world (Jian et al., 1996). These trends have been reinforced by the increasing territorial competition among Chinese provinces whereby politically stronger and economically wealthier local authorities have actively promoted the concentration of innovative activities at the expense of neighbouring (competing) areas (Chien and Gordon, 2008). In addition, internal restrictions on labour and capital mobility have played an important role in the development of the Chinese spatial economy (Duflo, 2010) and, as a consequence, on its geography of innovation Although restrictions on capital mobility have been progressively eased over time, enduring restrictions on labour movement in combinations with high levels of informal rural-urban migration have contributed to concentrate human and physical capital in urban areas, accelerating the prominence of urban centres as innovation hubs. Hence, in this first big phase of Chinese innovation development, the paradox is that we are seeing an unusually pure case of marketdriven agglomeration effects and the mutually supportive relationship of urbanization, localization and innovation, which owes much to a planned economy India India appears to have a radically different geography of innovation from China (Table 3a). R&D expenditures, in regressions (1) to (3), suggest a more conventional relationship between R&D inputs and innovation outputs. Unlike China, we find that regional R&D spending is important for regional innovation, with point estimates which are very large, although only significant at the 10% level. Moreover, R&D spending as a determinant of regional innovative capacity maintains its importance as social conditions and structural factors are introduced in the analysis. Unlike China, spillovers 22